The Genom of Homo sapiens.pdf

406 CHEUNG ET AL.Figure 3. The expression levels measured by microarrays foreight highly variable and two relatively nonvariable genes in 50individuals. Each point represents the expression level for an individual.of expression levels among the individuals studied wasfrom 14- to 48-fold. The expression levels of 8 of thehighly variable genes and 2 relatively nonvariable genesfor the 50 unrelated subjects are shown in Figure 3.EVIDENCE FOR HERITABILITYWe assessed the heritability of expression variation intwo ways. In the first method, we used data from sets ofindividuals with different degrees of relatedness, and inthe second, we used the resemblance between offspringand parent. Figure 4 shows the results from the first approachand compares the variance in expression levelamong three groups of subjects: 50 unrelated members ofCEPH families (parents of the large CEPH sibships), 10sets of siblings (also CEPH sibships), and 10 pairs of MZtwins. We found that the variance is largest among the unrelatedindividuals (a sample from the European population)and smallest between members of the MZ twin pairs.For all 5 genes shown in Figure 4 (and for 10 others thatwe have tested), we found the following pattern: as the degreeof relatedness increases, the variance in expressionlevels decreases, suggesting that genetic (germ-line) differencescontribute to variation in gene expression.Although this kind of comparison of variances providesa valid initial assessment of evidence for heritabilityin some sense, other, more standard methods have theadvantage of providing numerical estimates of the conventionalmeasure, the so-called “narrow-sense” heritability.In addition to data for the CEPH parents, we haveobtained gene expression data for their parents (CEPHgrandparents). It is known that the regression of offspringon mid-parent value (in standard linear regression) providesan estimate of the desired heritability (Falconer andMacKay 1996), so we used the expression data for the unrelatedCEPH parents and their parents to estimate this regressioncoefficient. A striking example is shown for glutathioneS-transferase M2, GSTM2, in Figure 5. We haveestimated the regression of parent on mid-grandparentthis way for all ~3800 genes on the microarray that wereexpressed in the lymphoblastoid cells. Approximately50% of these values are negative, and thus give no evidencefor a heritable component. Viewing the collectionas a whole, if no genes showed evidence for heritability,we would expect that the distribution of regression estimateswould have as many negative as positive values.Instead, we find an excess of large positive values. Theregression coefficient (b) is greater than 0.5 for 94 genes,but b < – 0.5 for only 48. Similarly, we find b > 0.75 for9 genes, but b < – 0.75 for only 4 genes.Of course, we consider these estimates at best crude indicatorsof heritability, especially since we have looked atso many genes. However, our interest in the estimate ofheritability is not mainly as an end in itself; we use it inconjunction with the variance ratio to select genes for furtherstudy. In the follow-up studies, usually by RT-PCR,we will carry out group comparisons (sibships, twins,etc.) like those shown in Figure 4, tests of closely linkedFigure 4. Variance in expression level for five genes. QuantitativeRT-PCR data for 50 unrelated individuals, 89 offspring in10 CEPH sibships, and 10 sets of monozygotic twins.Figure 5. Regression plot of expression level of GSTM2 for 50offspring–midparent pairs. The slope (represented by solid line)for the regression is 0.75 with a standard error of 0.45.